Phase 04 — Prompts and instructions
Part of Harness engineering.
Three layers of text steer an agent. The system prompt sets its role and rules. Memory files carry project facts into every session. Output contracts say what a reply must look like. This phase builds each layer, lints it, and shows where a prompt must hand off to code that enforces the rule.
Lessons
- System prompt and steering — five sections in a stable order, steering lines, and a guard that enforces what the prompt only asks.
- Memory files — keep CLAUDE.md and AGENTS.md short, checkable and pointer-based.
- Output contracts — extract, validate and re-prompt, with a retry bound.
Where the depth lives
| If you want | Read |
|---|---|
| The anatomy of a prompt at PM depth | The anatomy of a prompt |
| Prompting inside coding agents | Prompting inside coding agents |
| Structured output patterns | Structured prompting |
| Why prompts fail | When prompts fail |
| The principles behind a lean memory file | Ten principles of a working harness |
Test yourself
- Why must volatile data such as today’s date stay out of the system prompt? <details><summary>Answer</summary>The system prompt is the cached prefix. One changed byte early in it invalidates the cache for everything after it, so per-turn facts belong in the user message. (<a href=“./01-system-prompt-and-steering/docs/en.md”>Lesson 1</a>)</details>
- Your prompt says “Never edit .env”, yet the agent edits it once. What was wrong with the design? <details><summary>Answer</summary>A prompt only asks, and models can fail to comply. A rule that must hold needs enforcing code, such as a hook that denies the edit, with the prompt line as the polite explanation. (<a href=“./01-system-prompt-and-steering/docs/en.md”>Lesson 1</a>)</details>
- Where do a tone rule, an ask-before-deleting rule and a refusal style each belong? <details><summary>Answer</summary>Tone belongs in the output contract, ask-before-deleting in the workflow, and the refusal style in the constraints. Each steering line goes in the section it shapes. (<a href=“./01-system-prompt-and-steering/docs/en.md”>Lesson 1</a>)</details>
- Moving rules behind an @import keeps CLAUDE.md short. Does it save context? <details><summary>Answer</summary>No. Imported files load at launch with the file that imports them, so the cost stays. To save context, move path-specific rules to scoped rule files or use a skill that loads on demand. (<a href=“./02-memory-files/docs/en.md”>Lesson 2</a>)</details>
- The agent keeps repeating the same mistake. Why not just add a paragraph to CLAUDE.md? <details><summary>Answer</summary>Longer files lower adherence and cost budget every turn. Fix the harness instead, with a lint rule or a hook, and keep the file to facts that hold in every session. (<a href=“./02-memory-files/docs/en.md”>Lesson 2</a>)</details>
- A model returns valid JSON wrapped in prose with a trailing comma. What does the harness do, and what stops an endless loop? <details><summary>Answer</summary>It extracts the block, repairs the comma, then validates keys and types. When a reply still fails, it re-prompts with the exact violation, and a retry budget makes it fail loudly after a few attempts. (<a href=“./03-output-contracts/docs/en.md”>Lesson 3</a>)</details>
- When do you rely on the prompt contract alone, and when do you add a check or a schema? <details><summary>Answer</summary>A human reader can tolerate small drift, so the prompt is enough. Machine-read output needs a check, and a JSON schema through structured outputs where the API supports it. (<a href=“./03-output-contracts/docs/en.md”>Lesson 3</a>)</details>